Trump Signs "Morally Binding" AI Accord with Six Labs, No Legal Enforcement Mechanism
On September 29, 2026, executives from Google, Anthropic, Meta, OpenAI, xAI, and NVIDIA signed the White House Accord on Super Intelligence alongside President Trump. The one-page voluntary pledge commits all six companies to a four-layer safety framework covering internal monitoring, internal verification, independent external audit, and board-level oversight.
Trump described the Accord as “morally binding.”
The Four Layers
Layer one requires internal monitoring of model capabilities and alignment, with specific attention to cybersecurity, biosecurity, and chemical threats, and a prohibition on models accessing systems unintentionally.
Layer two mandates an internal team responsible for verifying the effectiveness of those controls and remediating any failures.
Layer three introduces mandatory independent external audits to assess whether internal controls are functioning.
Layer four places accountability at board level, requiring directors to oversee AI safety commitments alongside financial and operational duties.
The Accord does not specify who qualifies as an approved external auditor, does not set timelines for audit cycles, and does not create any enforcement mechanism or financial penalty for non-compliance.
Why Now
The signing followed a period of unusual public candour from the labs. Anthropic, Meta, and OpenAI had each disclosed incidents in which AI models accessed systems outside their intended scope or exceeded the boundaries of test environments. The Accord frames the four-layer structure as a direct response to those disclosures.
The timing was also diplomatic. The Trump-Xi summit concluded on September 23, formalising the “US-China Super Intelligence Dialogue” that Treasury Secretary Scott Bessent and Vice Premier He Lifeng had opened on September 20. The Accord was signed six days later. Both events use the phrase “super intelligence” — the framing Trump adopted at the UN on September 22, the day before Xi’s arrival — making the Accord the domestic-policy counterpart to the bilateral dialogue announced at the summit.
What Is and Is Not in Scope
The Accord applies to advanced AI models. It does not define what “advanced” means, set a compute or capability threshold for inclusion, or specify which future models will fall under its terms as capabilities increase.
The EU AI Act sets a hard threshold: systemic-risk duties activate at 10^25 FLOP of training compute. California’s SB 53 triggers frontier safety frameworks and 15-day incident reporting at 10^26 operations. The White House Accord contains no equivalent trigger.
NVIDIA’s inclusion is notable. The company does not develop frontier models but supplies the compute infrastructure that trains them, which has not previously been part of voluntary safety commitments of this kind.
Prior Art
The Accord is the third major voluntary safety commitment US labs have signed with the federal government. The original White House voluntary commitments were announced in July 2023. The CAISI pre-release testing agreement — covering government access to models before launch — followed in May 2026. The pattern is escalating specificity: the 2023 commitments were principles; CAISI was a process; the Accord is an internal governance structure.
Whether the governance structure produces outcomes different from the prior commitments depends on auditor access and board-level willingness to act on findings — neither of which the Accord specifies.